Pay20Y
我也遇到了这个问题，训练损失降得还可以，但是测试的准确率很低。目前训练识别分支时是gt标注的坐标进行RoI Rotate，并没有将检测的结果用来生成RoI，请问是否和这种训练方式有关？谢谢！ @btdaxian @ @jiangxiluning
recognizer has already been implemented. change the branch to develop
Sorry it's feature/reg branch
> > > How do you find it? I cannot find the recognizer code. use git branch and git checkout
I have met the same problem, how to solve it? Thanks!
您好，请问utils/recognition_evaluation.py具体应该怎么修改呢，我修改了_normalize_text这里，除此之外还需要修改什么呢？谢谢！
> > 您好，请问utils/recognition_evaluation.py具体应该怎么修改呢，我修改了_normalize_text这里，除此之外还需要修改什么呢？谢谢！ > > @Pay20Y 去掉gt中之前加入的分隔符 好的，感谢
The result is same, since affine transformation is essentially a matrix operation.
Sorry for waiting, maybe after Qingming vacation
Sorry, the epochs can not be specified directlly in this code, please calculate the corresponding number of iterations.